Transmeta: TM8000: Astro: Efficeon [Last updated: October 30, 2005]
Transmeta was founded in 1995 and ceased operations in 2009. Unlike most technology companies, the firm was founded upon the principle of power consumption and intended to introduce the most energy-efficient (“green”) microprocessors available to the market. This fundamental difference between other chip producers, who concentrated only on power output, turned enough heads that some manufacturers concerned with efficiency, heat production, and other issues on top of mere computing power turned to Transmeta. Unfortunately, the Transmeta chips’ computing power fell short of many expectations despite keeping in line with efficiency parameters. Eventually, the bigger competitors (Intel especially) noticed and followed suit in creating their own “greener” microprocessor versions. In fact, Intel admitted that were it not for Transmeta’s efforts, Intel would not have been pressured to develop the more energy-efficient Pentium M (mobile Centrino bundle).
This page is no longer being actively updated and is stored for archival purposes only. Many outgoing links to external domains are probably broken. You can find the complete history and latest information about Transmeta and its processors on Wikipedia (Transmeta).
Some of us are very interested in how successful Transmeta can be against Intel in the mobile (notebook) market. While the Crusoe offered new technology and some advantage in battery life, at the same time consumers had to be willing to give up performance for that advantage. During the time Crusoe has been out in the market, it must have been obvious to Transmeta that the effiency savings alone could not save it, especially when Intel began to shoulder its way into the ultramobile market with the introduction of the Pentium-M and Centrino package, a higher-powered CPU that also promised greater energy efficiency. So, after long years of development, the TM8000 “Astro,” officially named Efficeon, was released mid-October, 2003.
I’ve put together this no-longer-small page using material from Efficeon publications and subsequent articles related to it. The full Microprocessor Forum webcast presentation (slides) is available in PDF format from Transmeta’s news page, over 70 pages and over 7 megs (recommend a download and read for much more technical information including LongRun2, which I won’t discuss here). The following is just a small portion directly related to what I and many others are really interested in – performance and energy efficiency.
NOTES: I didn’t put much (if any) actual analysis on this page due to analyses on other sites (e.g. this saved Leog.net thread).
[The following information has not been updated since about 7/2004]
Products announcement list:
- IBASE MB860 Mini-ITX motherboard (OEM) (Efficeon-based mini-ITX form factor motherboard, available only to OEMs, as far as I know) : press 07/01/2004 (from June 2004 Computex in Taiwan), $?.
- Sharp Mebius Muramasa PC-MM50F (Efficeon subnote, upgrade to the PC-MM2 below) : press 05/12/2004, due June 11, 2004, 200000 Yen.
- Sharp Mebius Muramasa PC-CV50 (F or FW) (Efficeon mininotebook, smallest in the Mebius series from Sharp-JP) : press 05/12/2004, due June 18, 2004, 190000 Yen.
- Sharp Actius MM20 (Efficeon subnotebook, same as the Muramasa PC-MM2 below) : press 03/15/2004, April 2004, $1499.
- Hewlett Packard BX1000 Blade PC (Efficeon Blade PC and Crusoe Thin Clients) : press 12/04/2003, 3-4/2004. (Official Specifications)
- SHARP Mebius Muramasa PC-MM2 (Efficeon subnotebook) : Japanese press 12/08/03, Selling in Japan. MM2 for US above. (Japanese link) (Specifications, 5NE and 1DA models) (Japanese review with some tests)
Longrun 2 information:
- NEC: [03/25/04] “The Japanese computer maker has licensed Transmeta’s LongRun2 technologies for use throughout its 90-, 65- and 45-nanometer generation semiconductor products. NEC said it also purchased a small percentage of Transmeta’s common stock, giving it some control in the direction the Santa Clara, Calif.-based chipmaker turns in the future. Specific financial details were not available.” InternetNews
Sharp (Japan) had released at least one another TM8000 model since this page was last updated. The PC-MPxxx models are fairly interesting.
The OQO is also no longer vaporware, though it uses the Crusoe chip and not a TM8000. It’s only gotten lukewarm reviews for a while now because few people want such a super-miniscule computer at this point.
The Flybook, which had been whispered about for quite some time, is apparently going to be available in the US. However, for whatever reason that also has only a Crusoe chip and not an Efficeon (costs?), assuming article specs are correct. It’s a convertible tablet/subnotebook with a sunlight-readable screen.
Transmeta has since changed its business model to licensing technologies and intellectual property instead of actively developing and fabbing chips.
1) Efficeon Comparative Benchmarks and Roadmaps (Official)
The following “official” benchmarks from Transmeta compare the Efficeon to other chips such as the Pentium-M (or “Centrino” with the wireless package) or Pentium 4. I’ve either removed or condensed my benchmark commentary a great deal to make room for 3rd-party articles, which are more interesting anyway.
You’ll notice Transmeta differentiates some of the benchmarks at the magical “7 watt” limit, which is supposedly the maximum power point where a laptop can work without a fan. They explain that [here (Comparing Mobile Performance)]. Remember that “official” benchmarks from the product manufacturer are never as good as truely neutral and properly scientific third-party tests.
- Integer benchmarks (encryption tests):
- Floating point benchmarks (MFLOP = megaflop = one million floating-point operations per second; floating point = decimal-place number):
- 3D graphics benchmarks:
- The 3D graphics performance can be misleading; some will complain that it’s comparing apples to oranges, or that you’re not comparing the CPUs at all (which is true…). But I suppose Transmeta’s point here is that the TM8000 will outperform Intel when each have their own “native” graphics subset. Yes, Transmeta is using Nvidia here, but considering the graphics subsystems are included in the price of their respective CPUs, that would mean better marketability for a base TM8000 chip. Anandtech has a nice article talking about the Efficeon/nVidia combination and potential market power. A little light digging, however, suggests that the old, plain “3DMark 2001” benchmark may have had a bug that gave an advantage to Nvidia systems, but we don’t know what version 3DMark Transmeta decided to run. I hope it was the later SE version (currently the “pro” version)…
- [3D Graphics Performance (3DMark 2001)]
- System suite benchmarks:
The tests they used (commercial pages open up in a new window):
CPUMark99, MobileMark 2002, SysMark 2002, PCMark2002,
and Business Winstone 2001.
Notice that some of the benchmarks are available in newer iterations; Business Winstone,
for instance, is currently at 2002. The latest PCMark appears to be a “pro” version.
- [PC System Performance (at 7 Watt TDP MHz)]
- The tests they used (commercial pages open up in a new window):
- Energy Efficiency:
- Transmeta didn’t have actual efficiency comparison tests in their official TM8000 release, so they discussed idle CPU power consumption. While it could make a real-world difference in battery life, one must remember that laptops are by nature chimeric beasts, and many other factors go into determining overall energy efficiency of a system. (Discussed later in a 3rd-party article)
- [Efficeon consumes ~8x Lower Idle Power…]
- Please see the Roadmap section below for an interesting energy vs. speed roadmap comparison between Transmeta and Intel from a Japanese site.
- Small package designs = small consumer products:
- Not related to performance, this slide illustrates Efficeon’s ability to fit in small spaces, a high plus for ultramobile systems. This slide illustrates the TM8000 chip integrating Northbridge while Intel’s P-M has a separate Northbridge (area comparison).
- [Small Package Solutions]
- Efficeon Roadmaps:
- [TM8000 official roadmap]
- I’m adding here an interesting slide from a Japanese article in PC Watch that discusses the development of Efficeon. This slide was created by Hiroshige Goto, the author, and better illustrates the expected roadmaps of Efficeon generations 1 and 2 versus Intel’s (Pentium-M) Banias and Dothan cores on power (Watt) v. speed (GHz) axes.
2) Article Summary: Transmeta and VIA in competition with Intel: 12/12/2003
The Japanese site PC Watch posted an article (in Japanese, of course) a couple days ago discussing the development of Transmeta and VIA chips in competition with Intel’s Pentium-M. Since I’m not a native speaker of Japanese, and I’m sure most readers who happen on this page aren’t either, we’ll have to rely on the interesting machine-translated version produced by any one of the free online translators (Japanese is a very difficult language to translate via machine).
What I could get out of this (“*” comments are my own and not in the article):
Earlier in ’03, at the IDF Japan Spring conference, Japanese OEM manufacturers expressed interest in smaller packages. According to the Intel rep (Singer?), the current state of chip manufacturing was essentially the limit (“sufficient”) due to difficulty in laying the traces on motherboards of this size (*we’re seeing the technology go from 130nm to 90nm fairly soon, though). But the opinion of Japanese OEMs differed; one PC maker noted
that if a smaller CPU package is made, a smaller motherboard would also probably follow. Basically, it seems Intel’s stance at this conference was that the limitation was based on motherboard size, while the Japanese OEMs don’t see it as a legitimate obstacle in developing smaller machines. The next section of the article basically summarizes the difference between Japan and America in viewing smaller machines (*Oh so true…); the difficulty for Japan when talking about Intel chips is that the American laptop market does not currently favor “smaller is better” designs, which is exactly what Japan wants, but their market is much smaller than the US (and Intel is a US company blah blah). The essential trade-off is elementary economics: a smaller CPU/motherboard package will cost more, and higher cost offsets consumer purchases.
The next section of the article then discusses how Transmeta and VIA are directly competing against Intel for the Japanese appetite for small machines. Specifically, the package size between Transmeta’s Efficeon is compared to Centrino at 3,027 mm^2 and 3,592 mm^2 respectively (*looking at the figure in the article, the numbers differ a bit from the slide provided by Transmeta). Efficeon’s smaller overall package is obviously more favorable for smaller machines, along with the fact that they will move to a smaller 90nm process eventually, but it is currently limited by the fact that the Centrino package has a fully integrated graphics subset while the Efficeon package is not fully integrated with the NVidia solution.
The article then moves on to discuss a third competitor – one even more esoteric than Transmeta in the US – VIA Technologies. The C3/Eden/Antaur from VIA is an even smaller package than the Efficeon (*though I have no idea how it performs), and VIA is already working on Nano BGA technology to further reduce package size (September COMPUTEX Taipei). The motherboard developed for this new chip will be smaller than the Mini ITX form factor, called Nano ITX. Finally, the article concludes with the topic on absorbing cost increases while moving to smaller and smaller motherboards/chip packages. The article doesn’t seem to conceal hopes that the small-package/energy efficient designs from Transmeta/VIA will “destroy the wall of Intel” (*or that’s a bad machine translation…), and that more mini-note laptops and SFF (small form factor) PCs will be produced.
3) Article Summary: Comparing Efficeon efficiency to ultralow voltage edition Pentium-M laptops: 12/17/2003
This is a very interesting article albeit a little difficult to slosh through without a good understanding of Japanese (PC Watch). At least the charts and tables are straightforward, with a little help from the Babelfish. I’m changing the Article Summary format to include my own commentary as an interpretation of the (interpreted) article since it gets messy to continually annotate.
The article begins with an introduction to the engineering advantages of the Efficeon over its predecessor, the Crusoe. Of interest is the supposed energy efficiency of Efficeon-based systems over Pentium-M-based systems. The test systems include the Sony X505 Extreme, Sony TR (“Bio note”?), Loox T (also known as the Fujitsu P-5000), JVC Interlink XP7, Panasonic CF-R2, and the Sharp PC-MM2-5NE (the only Efficeon system since that’s the only Efficeon-based laptop to be released within the short term). Importantly, battery capacities are listed nicely in a table.
PC Watch used the MobileMark2002 benchmarking software to test these laptops. It’s important to note that this benchmarking suite measures performance and overall battery life, but does not in itself take into account the actual battery capacity, which is a very critical factor for determining actual efficiency. Graph 2 is therefore misleading without knowing additional details. It’s easy to add an hour or two just by having a larger battery – mediocre reviews that compare the Fujitsu P-2000 and P-5000 series fail to note that the P-5000 series has a much larger battery to begin with. Hence the battery capacity table in watt-hours provided by PC Watch is very useful. The other tables, charts, and benchmarks are relatively easy to figure out.
Graph 3 is Average electric power consumption (battery capacity / drive time) unit in watts. This is probably the most important chart in terms of overall system efficiency; PC Watch divided battery capacity by “drive time” (i.e. running time). I want to stress that it’simpossible to say “Efficeon is more/less energy efficient than the Pentium-M” in this case because these are complete laptop systems being compared, not merely the CPU. Laptops being chimeric beasts; one must also consider the efficiency of the hard drive, LCD, type and quantity of RAM, video, and other factors.
Graph 4 is the throughput per 1W (MobileMark2002 Performance Rating / average electric power consumption), or as I interpret it, actual performance per watt power.
Graphs 5-9 go over productivity benchmarks which I won’t reproduce or discuss here (visit the original article as the numerical charts are very straightforward). But it is interesting to note that in most of the laptop system benchmarks, the Efficeon-based MM2 lagged behind other Pentium-M systems.
“Chart 2” (the really big table) lists a number of processes and time measurements for the Sony PCG-X505, Sharp MM1, and MM2. (*)
The article section concerning the 1.8″ HDD performance compared to 2.5″ merely states that the performance is roughly the same in the drives they compared (HGST DK14FA-20, MK2004GAL, HTS548080M9AT00, IC25N080ATMR04-0).
The conclusion seems to say that the TM8600-1GHz in the Sharp MM2 appears inferior to the Pentium-M 1GHz, while it’s about par with the 900MHz (this is in-line with the official benchmarks, and likely PC Watch is speaking on average of the P-M systems as opposed to direct CPU measurement). Someone using a Crusoe would definitely notice a difference, according to the article; I’m not positive, but the article then seems to state the operating difference (opening windows?) between the MM2 and the P-M systems was not noticable. The highest grade clock for the TM8000 processor, 1.1 GHz, isn’t yet used in a system, so PC Watch comments that this clock will probably compete with the ultralow voltage edition P-M 1GHz. The article concludes with a general call for more vendors to make Efficeon-based PCs.
4) Article Summary: Special Edition report on Efficeon (CMS): 01/13/2004
By Satoru Yoneda, this technically detailed review at PC Web focuses on benchmarking three low-voltage laptops at the time. I actually had this site bookmarked for a while but never got around to adding it. My bad.
Basically, the reviewer had gotten one of the first Sharp MM2 laptops (PC-MM2-5NE) and compared it against the old Sharp MM1 (Crusoe) and the Panasonic CF-R2 (Pentium-M 1GHz). He does state that the MM2 model he tested was a demo, and could differ from actual models. The attention is essentially given to Efficeon’s CMS (Code Morphing Software), the solution intended to supply both energy efficiency and high performance. The entire review is 10 site pages long in Japanese, so instead of going over each page (which
English-readers can do with translators such as Babelfish), I’ll try to summarize the conclusions.
In measuring power efficiency, the reviewer removed the batteries and measured the electric current of AC adapter output in real time (volts, to measure the power consumption of the PC itself directly). Because the “power-saving” technologies (LongRun and SpeedStep) shouldn’t be active at full AC power, doing so would provide a baseline comparison for actual efficiency on battery (*my interpretation). In one test, the reviewer commented that the Transmeta Efficeon’s supposed lower power consumption in Windows XP idle state was exaggerated compared to the other two chips. Satoru then goes on to discuss TDP (Thermal Design Power), which was a major marketing point for the Efficeon.
Unfortunately the rough translation is a bit difficult to follow; it seems the Crusoe
(and Efficeon, according to the graphs) consumes 18W when Windows XP boots. The TDP of the Crusoe machine is 6.5W, and Efficeon at 1GHz is 5W. However, I don’t understand what
the author means in commenting about heat output and the 8W figure. It would be nice to
understand some technical Japanese. 😛
In the Panasonic R2, “rise of electric power consumption stopped in barely 3W” and the TDP of the ultralow voltage edition Pentium M was measured at 7W. Satoru makes additional commentary for the Pentium M machine, but it’s difficult to follow in rough translation.
The last graph for power consumption concerns “waiting electrical power” – I assume this is stand-by mode. The MM2 had the lowest power intake at this level, around 200mW, lower than the R2. Satoru assumes the 1-week stand-by assertion can be proven by the low power intake of the Efficeon in this state.
5) Sharp Actius MM20 reviews in English (some of these are more like spec blurbs than a real review)
LAPTOP Magazine 04/19/04:
“..[T]he Actius MM20, was rated a “Best Buy” by LAPTOP magazine in the April 2004 issue.” No online article available.
TIME Magazine 04/14/04:
This review is typical of all such flawed consumer reviews. “Also, I’ve seen longer battery life in chunkier notebooks with Intel’s Centrino chipset.” – well DUH, that’s why they’re chunkier. Such a silly thing to write. In any case, it offers a very basic opinion about the MM20, but in terms of actual content we can’t get from mere specifications, there isn’t any.
Van’s Hardware 04/04/04:
This is a particularly critical English article about the MM20 which benchmarks the Efficeon against several other CPUs. It definitely gives pause compared to the other benchmarks available, but my main gripe with this technical article is that it doesn’t look at power efficiency at all, which is the entire selling point of the Efficeon; Transmeta would have died years ago if pure performance was the only issue. You can’t compare an Efficeon with an Athlon 64 or P-4 no matter which way you cut it. As far as I’m concerned, this article compared the wrong things.
New York Times 03/25/04:
No on-line article, this was published and available for
download-purchase only. I haven’t actually read it.
PC Magazine 03/15/04:
A much more academic consumer review, and includes some benchmarking. “This score is commendable considering the battery’s small size (1.8 Ahrs).” – this tells me the reviewer is certainly able to look past the chassis when it comes to performance marks, unlike many other consumer reviews that fail to even compare baseline battery capacities. The benchmarks against its main market rival (pentium-M) are interesting as well – it just doesn’t fare as well, but whether that is from other component bottlenecks, I don’t know (e.g. different RAM capacities).
Pretty much just a spec blurb.